Google Drops Gemini Decision 1

Google is turning model inference into a decision primitive, which could make agent pipelines faster and dramatically cheaper.

• –Supports up to 1,000 questions against one shared multimodal input.
• –Returns probabilities and confidence scores instead of verbose generated text.
• –Zero output-token billing changes the economics of routing, filtering, and verification.
• –The real test is calibration: independent benchmarks must show whether confidence tracks correctness.
• –Developers should compare it against lightweight classifiers and smaller LLMs on latency, cost, and reliability.